libc: Attempt to execute new atexit() destructors during __cxa_finalize().
[dragonfly.git] / sys / vfs / hammer / hammer_object.c
1 /*
2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34
35 #include "hammer.h"
36
37 static int hammer_mem_lookup(hammer_cursor_t cursor);
38 static int hammer_mem_first(hammer_cursor_t cursor);
39 static int hammer_frontend_trunc_callback(hammer_record_t record,
40                                 void *data __unused);
41 static int hammer_bulk_scan_callback(hammer_record_t record, void *data);
42 static int hammer_record_needs_overwrite_delete(hammer_record_t record);
43 static int hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
44                                 hammer_btree_leaf_elm_t leaf);
45 static int hammer_cursor_localize_data(hammer_mount_t hmp, hammer_data_ondisk_t data,
46                                 hammer_btree_leaf_elm_t leaf);
47
48 struct rec_trunc_info {
49         uint16_t        rec_type;
50         int64_t         trunc_off;
51 };
52
53 struct hammer_bulk_info {
54         hammer_record_t record;
55         hammer_record_t conflict;
56 };
57
58 /*
59  * Red-black tree support.  Comparison code for insertion.
60  */
61 static int
62 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
63 {
64         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
65                 return(-1);
66         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
67                 return(1);
68
69         if (rec1->leaf.base.key < rec2->leaf.base.key)
70                 return(-1);
71         if (rec1->leaf.base.key > rec2->leaf.base.key)
72                 return(1);
73
74         /*
75          * For search & insertion purposes records deleted by the
76          * frontend or deleted/committed by the backend are silently
77          * ignored.  Otherwise pipelined insertions will get messed
78          * up.
79          *
80          * rec1 is greater then rec2 if rec1 is marked deleted.
81          * rec1 is less then rec2 if rec2 is marked deleted.
82          *
83          * Multiple deleted records may be present, do not return 0
84          * if both are marked deleted.
85          */
86         if (rec1->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
87                            HAMMER_RECF_COMMITTED)) {
88                 return(1);
89         }
90         if (rec2->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
91                            HAMMER_RECF_COMMITTED)) {
92                 return(-1);
93         }
94
95         return(0);
96 }
97
98 /*
99  * Basic record comparison code similar to hammer_btree_cmp().
100  *
101  * obj_id is not compared and may not yet be assigned in the record.
102  */
103 static int
104 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
105 {
106         if (elm->rec_type < rec->leaf.base.rec_type)
107                 return(-3);
108         if (elm->rec_type > rec->leaf.base.rec_type)
109                 return(3);
110
111         if (elm->key < rec->leaf.base.key)
112                 return(-2);
113         if (elm->key > rec->leaf.base.key)
114                 return(2);
115
116         /*
117          * Never match against an item deleted by the frontend
118          * or backend, or committed by the backend.
119          *
120          * elm is less then rec if rec is marked deleted.
121          */
122         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
123                           HAMMER_RECF_COMMITTED)) {
124                 return(-1);
125         }
126         return(0);
127 }
128
129 /*
130  * Ranged scan to locate overlapping record(s).  This is used by
131  * hammer_ip_get_bulk() to locate an overlapping record.  We have
132  * to use a ranged scan because the keys for data records with the
133  * same file base offset can be different due to differing data_len's.
134  *
135  * NOTE: The base file offset of a data record is (key - data_len), not (key).
136  */
137 static int
138 hammer_rec_overlap_cmp(hammer_record_t rec, void *data)
139 {
140         struct hammer_bulk_info *info = data;
141         hammer_btree_leaf_elm_t leaf = &info->record->leaf;
142
143         if (rec->leaf.base.rec_type < leaf->base.rec_type)
144                 return(-3);
145         if (rec->leaf.base.rec_type > leaf->base.rec_type)
146                 return(3);
147
148         /*
149          * Overlap compare
150          */
151         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
152                 /* rec_beg >= leaf_end */
153                 if (rec->leaf.base.key - rec->leaf.data_len >= leaf->base.key)
154                         return(2);
155                 /* rec_end <= leaf_beg */
156                 if (rec->leaf.base.key <= leaf->base.key - leaf->data_len)
157                         return(-2);
158         } else {
159                 if (rec->leaf.base.key < leaf->base.key)
160                         return(-2);
161                 if (rec->leaf.base.key > leaf->base.key)
162                         return(2);
163         }
164
165         /*
166          * We have to return 0 at this point, even if DELETED_FE is set,
167          * because returning anything else will cause the scan to ignore
168          * one of the branches when we really want it to check both.
169          */
170         return(0);
171 }
172
173 /*
174  * RB_SCAN comparison code for hammer_mem_first().  The argument order
175  * is reversed so the comparison result has to be negated.  key_beg and
176  * key_end are both range-inclusive.
177  *
178  * Localized deletions are not cached in-memory.
179  */
180 static
181 int
182 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
183 {
184         hammer_cursor_t cursor = data;
185         int r;
186
187         r = hammer_rec_cmp(&cursor->key_beg, rec);
188         if (r > 1)
189                 return(-1);
190         r = hammer_rec_cmp(&cursor->key_end, rec);
191         if (r < -1)
192                 return(1);
193         return(0);
194 }
195
196 /*
197  * This compare function is used when simply looking up key_beg.
198  */
199 static
200 int
201 hammer_rec_find_cmp(hammer_record_t rec, void *data)
202 {
203         hammer_cursor_t cursor = data;
204         int r;
205
206         r = hammer_rec_cmp(&cursor->key_beg, rec);
207         if (r > 1)
208                 return(-1);
209         if (r < -1)
210                 return(1);
211         return(0);
212 }
213
214 /*
215  * Locate blocks within the truncation range.  Partial blocks do not count.
216  */
217 static
218 int
219 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
220 {
221         struct rec_trunc_info *info = data;
222
223         if (rec->leaf.base.rec_type < info->rec_type)
224                 return(-1);
225         if (rec->leaf.base.rec_type > info->rec_type)
226                 return(1);
227
228         switch(rec->leaf.base.rec_type) {
229         case HAMMER_RECTYPE_DB:
230                 /*
231                  * DB record key is not beyond the truncation point, retain.
232                  */
233                 if (rec->leaf.base.key < info->trunc_off)
234                         return(-1);
235                 break;
236         case HAMMER_RECTYPE_DATA:
237                 /*
238                  * DATA record offset start is not beyond the truncation point,
239                  * retain.
240                  */
241                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
242                         return(-1);
243                 break;
244         default:
245                 hpanic("unexpected record type");
246         }
247
248         /*
249          * The record start is >= the truncation point, return match,
250          * the record should be destroyed.
251          */
252         return(0);
253 }
254
255 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
256
257 /*
258  * Allocate a record for the caller to finish filling in.  The record is
259  * returned referenced.  In order to manually set data call this function
260  * with data_len=0 and then manually set record->leaf.data_len and
261  * record->data later.
262  */
263 hammer_record_t
264 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
265 {
266         hammer_record_t record;
267         hammer_mount_t hmp;
268
269         hmp = ip->hmp;
270         ++hammer_count_records;
271         record = kmalloc(sizeof(*record), hmp->m_misc,
272                          M_WAITOK | M_ZERO | M_USE_RESERVE);
273         record->flush_state = HAMMER_FST_IDLE;
274         record->ip = ip;
275         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
276         record->leaf.data_len = data_len;
277         hammer_ref(&record->lock);
278
279         if (data_len) {
280                 record->data = kmalloc(data_len, hmp->m_misc, M_WAITOK | M_ZERO);
281                 record->flags |= HAMMER_RECF_ALLOCDATA;
282                 ++hammer_count_record_datas;
283         }
284
285         return (record);
286 }
287
288 void
289 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
290 {
291         while (record->flush_state == HAMMER_FST_FLUSH) {
292                 record->flags |= HAMMER_RECF_WANTED;
293                 tsleep(record, 0, ident, 0);
294         }
295 }
296
297 /*
298  * Called from the backend, hammer_inode.c, after a record has been
299  * flushed to disk.  The record has been exclusively locked by the
300  * caller and interlocked with BE.
301  *
302  * We clean up the state, unlock, and release the record (the record
303  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
304  */
305 void
306 hammer_flush_record_done(hammer_record_t record, int error)
307 {
308         hammer_inode_t target_ip;
309
310         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
311         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
312
313         /*
314          * If an error occured, the backend was unable to sync the
315          * record to its media.  Leave the record intact.
316          */
317         if (error) {
318                 hammer_critical_error(record->ip->hmp, record->ip, error,
319                                       "while flushing record");
320         }
321
322         --record->flush_group->refs;
323         record->flush_group = NULL;
324
325         /*
326          * Adjust the flush state and dependancy based on success or
327          * failure.
328          */
329         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
330                 if ((target_ip = record->target_ip) != NULL) {
331                         TAILQ_REMOVE(&target_ip->target_list, record,
332                                      target_entry);
333                         record->target_ip = NULL;
334                         hammer_test_inode(target_ip);
335                 }
336                 record->flush_state = HAMMER_FST_IDLE;
337         } else {
338                 if (record->target_ip) {
339                         record->flush_state = HAMMER_FST_SETUP;
340                         hammer_test_inode(record->ip);
341                         hammer_test_inode(record->target_ip);
342                 } else {
343                         record->flush_state = HAMMER_FST_IDLE;
344                 }
345         }
346         record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
347
348         /*
349          * Cleanup
350          */
351         if (record->flags & HAMMER_RECF_WANTED) {
352                 record->flags &= ~HAMMER_RECF_WANTED;
353                 wakeup(record);
354         }
355         hammer_rel_mem_record(record);
356 }
357
358 /*
359  * Release a memory record.  Records marked for deletion are immediately
360  * removed from the RB-Tree but otherwise left intact until the last ref
361  * goes away.
362  */
363 void
364 hammer_rel_mem_record(hammer_record_t record)
365 {
366         hammer_mount_t hmp;
367         hammer_reserve_t resv;
368         hammer_inode_t ip;
369         hammer_inode_t target_ip;
370         int diddrop;
371
372         hammer_rel(&record->lock);
373
374         if (hammer_norefs(&record->lock)) {
375                 /*
376                  * Upon release of the last reference wakeup any waiters.
377                  * The record structure may get destroyed so callers will
378                  * loop up and do a relookup.
379                  *
380                  * WARNING!  Record must be removed from RB-TREE before we
381                  * might possibly block.  hammer_test_inode() can block!
382                  */
383                 ip = record->ip;
384                 hmp = ip->hmp;
385
386                 /*
387                  * Upon release of the last reference a record marked deleted
388                  * by the front or backend, or committed by the backend,
389                  * is destroyed.
390                  */
391                 if (record->flags & (HAMMER_RECF_DELETED_FE |
392                                      HAMMER_RECF_DELETED_BE |
393                                      HAMMER_RECF_COMMITTED)) {
394                         KKASSERT(hammer_isactive(&ip->lock) > 0);
395                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
396
397                         /*
398                          * target_ip may have zero refs, we have to ref it
399                          * to prevent it from being ripped out from under
400                          * us.
401                          */
402                         if ((target_ip = record->target_ip) != NULL) {
403                                 TAILQ_REMOVE(&target_ip->target_list,
404                                              record, target_entry);
405                                 record->target_ip = NULL;
406                                 hammer_ref(&target_ip->lock);
407                         }
408
409                         /*
410                          * Remove the record from the RB-Tree
411                          */
412                         if (record->flags & HAMMER_RECF_ONRBTREE) {
413                                 RB_REMOVE(hammer_rec_rb_tree,
414                                           &ip->rec_tree,
415                                           record);
416                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
417                                 KKASSERT(ip->rsv_recs > 0);
418                                 if (RB_EMPTY(&ip->rec_tree)) {
419                                         ip->flags &= ~HAMMER_INODE_XDIRTY;
420                                         ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
421                                 }
422                                 diddrop = 1;
423                         } else {
424                                 diddrop = 0;
425                         }
426
427                         /*
428                          * We must wait for any direct-IO to complete before
429                          * we can destroy the record because the bio may
430                          * have a reference to it.
431                          */
432                         if (record->gflags &
433                            (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL)) {
434                                 hammer_io_direct_wait(record);
435                         }
436
437                         /*
438                          * Account for the completion after the direct IO
439                          * has completed.
440                          */
441                         if (diddrop) {
442                                 --hmp->rsv_recs;
443                                 --ip->rsv_recs;
444                                 hmp->rsv_databytes -= record->leaf.data_len;
445
446                                 if (RB_EMPTY(&ip->rec_tree))
447                                         hammer_test_inode(ip);
448                                 if ((ip->flags & HAMMER_INODE_RECSW) &&
449                                     ip->rsv_recs <= hammer_limit_inode_recs/2) {
450                                         ip->flags &= ~HAMMER_INODE_RECSW;
451                                         wakeup(&ip->rsv_recs);
452                                 }
453                         }
454
455                         /*
456                          * Do this test after removing record from the RB-Tree.
457                          */
458                         if (target_ip) {
459                                 hammer_test_inode(target_ip);
460                                 hammer_rel_inode(target_ip, 0);
461                         }
462
463                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
464                                 --hammer_count_record_datas;
465                                 kfree(record->data, hmp->m_misc);
466                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
467                         }
468
469                         /*
470                          * Release the reservation.
471                          *
472                          * If the record was not committed we can theoretically
473                          * undo the reservation.  However, doing so might
474                          * create weird edge cases with the ordering of
475                          * direct writes because the related buffer cache
476                          * elements are per-vnode.  So we don't try.
477                          */
478                         if ((resv = record->resv) != NULL) {
479                                 /* XXX undo leaf.data_offset,leaf.data_len */
480                                 hammer_blockmap_reserve_complete(hmp, resv);
481                                 record->resv = NULL;
482                         }
483                         record->data = NULL;
484                         --hammer_count_records;
485                         kfree(record, hmp->m_misc);
486                 }
487         }
488 }
489
490 /*
491  * Record visibility depends on whether the record is being accessed by
492  * the backend or the frontend.  Backend tests ignore the frontend delete
493  * flag.  Frontend tests do NOT ignore the backend delete/commit flags and
494  * must also check for commit races.
495  *
496  * Return non-zero if the record is visible, zero if it isn't or if it is
497  * deleted.  Returns 0 if the record has been comitted (unless the special
498  * delete-visibility flag is set).  A committed record must be located
499  * via the media B-Tree.  Returns non-zero if the record is good.
500  *
501  * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory
502  * records to be returned.  This is so pending deletions are detected
503  * when using an iterator to locate an unused hash key, or when we need
504  * to locate historical records on-disk to destroy.
505  */
506 static __inline
507 int
508 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
509 {
510         if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY)
511                 return(1);
512         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
513                 if (record->flags & (HAMMER_RECF_DELETED_BE |
514                                      HAMMER_RECF_COMMITTED)) {
515                         return(0);
516                 }
517         } else {
518                 if (record->flags & (HAMMER_RECF_DELETED_FE |
519                                      HAMMER_RECF_DELETED_BE |
520                                      HAMMER_RECF_COMMITTED)) {
521                         return(0);
522                 }
523         }
524         return(1);
525 }
526
527 /*
528  * This callback is used as part of the RB_SCAN function for in-memory
529  * records.  We terminate it (return -1) as soon as we get a match.
530  *
531  * This routine is used by frontend code.
532  *
533  * The primary compare code does not account for ASOF lookups.  This
534  * code handles that case as well as a few others.
535  */
536 static
537 int
538 hammer_rec_scan_callback(hammer_record_t rec, void *data)
539 {
540         hammer_cursor_t cursor = data;
541
542         /*
543          * We terminate on success, so this should be NULL on entry.
544          */
545         KKASSERT(cursor->iprec == NULL);
546
547         /*
548          * Skip if the record was marked deleted or committed.
549          */
550         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
551                 return(0);
552
553         /*
554          * Skip if not visible due to our as-of TID
555          */
556         if (cursor->flags & HAMMER_CURSOR_ASOF) {
557                 if (cursor->asof < rec->leaf.base.create_tid)
558                         return(0);
559                 if (rec->leaf.base.delete_tid &&
560                     cursor->asof >= rec->leaf.base.delete_tid) {
561                         return(0);
562                 }
563         }
564
565         /*
566          * ref the record.  The record is protected from backend B-Tree
567          * interactions by virtue of the cursor's IP lock.
568          */
569         hammer_ref(&rec->lock);
570
571         /*
572          * The record may have been deleted or committed while we
573          * were blocked.  XXX remove?
574          */
575         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
576                 hammer_rel_mem_record(rec);
577                 return(0);
578         }
579
580         /*
581          * Set the matching record and stop the scan.
582          */
583         cursor->iprec = rec;
584         return(-1);
585 }
586
587
588 /*
589  * Lookup an in-memory record given the key specified in the cursor.  Works
590  * just like hammer_btree_lookup() but operates on an inode's in-memory
591  * record list.
592  *
593  * The lookup must fail if the record is marked for deferred deletion.
594  *
595  * The API for mem/btree_lookup() does not mess with the ATE/EOF bits.
596  */
597 static
598 int
599 hammer_mem_lookup(hammer_cursor_t cursor)
600 {
601         KKASSERT(cursor->ip != NULL);
602         if (cursor->iprec) {
603                 hammer_rel_mem_record(cursor->iprec);
604                 cursor->iprec = NULL;
605         }
606         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
607                                    hammer_rec_scan_callback, cursor);
608
609         return (cursor->iprec ? 0 : ENOENT);
610 }
611
612 /*
613  * hammer_mem_first() - locate the first in-memory record matching the
614  * cursor within the bounds of the key range.
615  *
616  * WARNING!  API is slightly different from btree_first().  hammer_mem_first()
617  * will set ATEMEM the same as MEMEOF, and does not return any error.
618  */
619 static
620 int
621 hammer_mem_first(hammer_cursor_t cursor)
622 {
623         KKASSERT(cursor->ip != NULL);
624         if (cursor->iprec) {
625                 hammer_rel_mem_record(cursor->iprec);
626                 cursor->iprec = NULL;
627         }
628         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_scan_cmp,
629                                    hammer_rec_scan_callback, cursor);
630
631         if (cursor->iprec)
632                 cursor->flags &= ~(HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM);
633         else
634                 cursor->flags |= HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM;
635
636         return (cursor->iprec ? 0 : ENOENT);
637 }
638
639 /************************************************************************
640  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
641  ************************************************************************
642  *
643  * These functions manipulate in-memory records.  Such records typically
644  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
645  */
646
647 /*
648  * Add a directory entry (dip,ncp) which references inode (ip).
649  *
650  * Note that the low 32 bits of the namekey are set temporarily to create
651  * a unique in-memory record, and may be modified a second time when the
652  * record is synchronized to disk.  In particular, the low 32 bits cannot be
653  * all 0's when synching to disk, which is not handled here.
654  *
655  * NOTE: bytes does not include any terminating \0 on name, and name might
656  * not be terminated.
657  */
658 int
659 hammer_ip_add_direntry(hammer_transaction_t trans,
660                      hammer_inode_t dip, const char *name, int bytes,
661                      hammer_inode_t ip)
662 {
663         struct hammer_cursor cursor;
664         hammer_record_t record;
665         int error;
666         uint32_t max_iterations;
667
668         KKASSERT(dip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY);
669
670         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
671
672         record->type = HAMMER_MEM_RECORD_ADD;
673         record->leaf.base.localization = dip->obj_localization |
674                                          hammer_dir_localization(dip);
675         record->leaf.base.obj_id = dip->obj_id;
676         record->leaf.base.key = hammer_direntry_namekey(dip, name, bytes,
677                                                          &max_iterations);
678         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
679         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
680         record->data->entry.obj_id = ip->obj_id;
681         record->data->entry.localization = ip->obj_localization;
682         bcopy(name, record->data->entry.name, bytes);
683
684         ++ip->ino_data.nlinks;
685         ip->ino_data.ctime = trans->time;
686         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
687
688         /*
689          * Find an unused namekey.  Both the in-memory record tree and
690          * the B-Tree are checked.  We do not want historically deleted
691          * names to create a collision as our iteration space may be limited,
692          * and since create_tid wouldn't match anyway an ASOF search
693          * must be used to locate collisions.
694          *
695          * delete-visibility is set so pending deletions do not give us
696          * a false-negative on our ability to use an iterator.
697          *
698          * The iterator must not rollover the key.  Directory keys only
699          * use the positive key space.
700          */
701         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
702         cursor.key_beg = record->leaf.base;
703         cursor.flags |= HAMMER_CURSOR_ASOF;
704         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
705         cursor.asof = ip->obj_asof;
706
707         while (hammer_ip_lookup(&cursor) == 0) {
708                 ++record->leaf.base.key;
709                 KKASSERT(record->leaf.base.key > 0);
710                 cursor.key_beg.key = record->leaf.base.key;
711                 if (--max_iterations == 0) {
712                         hammer_rel_mem_record(record);
713                         hmkprintf(trans->hmp, "Failed to find an unused namekey\n");
714                         error = ENOSPC;
715                         goto failed;
716                 }
717         }
718
719         /*
720          * The target inode and the directory entry are bound together.
721          */
722         record->target_ip = ip;
723         record->flush_state = HAMMER_FST_SETUP;
724         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
725
726         /*
727          * The inode now has a dependancy and must be taken out of the idle
728          * state.  An inode not in an idle state is given an extra reference.
729          *
730          * When transitioning to a SETUP state flag for an automatic reflush
731          * when the dependancies are disposed of if someone is waiting on
732          * the inode.
733          */
734         if (ip->flush_state == HAMMER_FST_IDLE) {
735                 hammer_ref(&ip->lock);
736                 ip->flush_state = HAMMER_FST_SETUP;
737                 if (ip->flags & HAMMER_INODE_FLUSHW)
738                         ip->flags |= HAMMER_INODE_REFLUSH;
739         }
740         error = hammer_mem_add(record);
741         if (error == 0) {
742                 dip->ino_data.mtime = trans->time;
743                 dip->ino_data.ctime = trans->time;
744                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY);
745         }
746 failed:
747         hammer_done_cursor(&cursor);
748         return(error);
749 }
750
751 /*
752  * Delete the directory entry and update the inode link count.  The
753  * cursor must be seeked to the directory entry record being deleted.
754  *
755  * The related inode should be share-locked by the caller.  The caller is
756  * on the frontend.  It could also be NULL indicating that the directory
757  * entry being removed has no related inode.
758  *
759  * This function can return EDEADLK requiring the caller to terminate
760  * the cursor, any locks, wait on the returned record, and retry.
761  */
762 int
763 hammer_ip_del_direntry(hammer_transaction_t trans,
764                      hammer_cursor_t cursor, hammer_inode_t dip,
765                      hammer_inode_t ip)
766 {
767         hammer_record_t record;
768         int error;
769
770         if (hammer_cursor_inmem(cursor)) {
771                 /*
772                  * In-memory (unsynchronized) records can simply be freed.
773                  *
774                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
775                  * by the backend, we must still avoid races against the
776                  * backend potentially syncing the record to the media.
777                  *
778                  * We cannot call hammer_ip_delete_record(), that routine may
779                  * only be called from the backend.
780                  */
781                 record = cursor->iprec;
782                 if (record->flags & (HAMMER_RECF_INTERLOCK_BE |
783                                      HAMMER_RECF_DELETED_BE |
784                                      HAMMER_RECF_COMMITTED)) {
785                         KKASSERT(cursor->deadlk_rec == NULL);
786                         hammer_ref(&record->lock);
787                         cursor->deadlk_rec = record;
788                         error = EDEADLK;
789                 } else {
790                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
791                         record->flags |= HAMMER_RECF_DELETED_FE;
792                         error = 0;
793                 }
794         } else {
795                 /*
796                  * If the record is on-disk we have to queue the deletion by
797                  * the record's key.  This also causes lookups to skip the
798                  * record (lookups for the purposes of finding an unused
799                  * directory key do not skip the record).
800                  */
801                 KKASSERT(dip->flags &
802                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
803                 record = hammer_alloc_mem_record(dip, 0);
804                 record->type = HAMMER_MEM_RECORD_DEL;
805                 record->leaf.base = cursor->leaf->base;
806                 KKASSERT(dip->obj_id == record->leaf.base.obj_id);
807
808                 /*
809                  * ip may be NULL, indicating the deletion of a directory
810                  * entry which has no related inode.
811                  */
812                 record->target_ip = ip;
813                 if (ip) {
814                         record->flush_state = HAMMER_FST_SETUP;
815                         TAILQ_INSERT_TAIL(&ip->target_list, record,
816                                           target_entry);
817                 } else {
818                         record->flush_state = HAMMER_FST_IDLE;
819                 }
820
821                 /*
822                  * The inode now has a dependancy and must be taken out of
823                  * the idle state.  An inode not in an idle state is given
824                  * an extra reference.
825                  *
826                  * When transitioning to a SETUP state flag for an automatic
827                  * reflush when the dependancies are disposed of if someone
828                  * is waiting on the inode.
829                  */
830                 if (ip && ip->flush_state == HAMMER_FST_IDLE) {
831                         hammer_ref(&ip->lock);
832                         ip->flush_state = HAMMER_FST_SETUP;
833                         if (ip->flags & HAMMER_INODE_FLUSHW)
834                                 ip->flags |= HAMMER_INODE_REFLUSH;
835                 }
836
837                 error = hammer_mem_add(record);
838         }
839
840         /*
841          * One less link.  The file may still be open in the OS even after
842          * all links have gone away.
843          *
844          * We have to terminate the cursor before syncing the inode to
845          * avoid deadlocking against ourselves.  XXX this may no longer
846          * be true.
847          *
848          * If nlinks drops to zero and the vnode is inactive (or there is
849          * no vnode), call hammer_inode_unloadable_check() to zonk the
850          * inode.  If we don't do this here the inode will not be destroyed
851          * on-media until we unmount.
852          */
853         if (error == 0) {
854                 if (ip) {
855                         --ip->ino_data.nlinks;  /* do before we might block */
856                         ip->ino_data.ctime = trans->time;
857                 }
858                 dip->ino_data.mtime = trans->time;
859                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
860                 if (ip) {
861                         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
862                         if (ip->ino_data.nlinks == 0 &&
863                             (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
864                                 hammer_done_cursor(cursor);
865                                 hammer_inode_unloadable_check(ip, 1);
866                                 hammer_flush_inode(ip, 0);
867                         }
868                 }
869
870         }
871         return(error);
872 }
873
874 /*
875  * Add a record to an inode.
876  *
877  * The caller must allocate the record with hammer_alloc_mem_record(ip,len) and
878  * initialize the following additional fields that are not initialized by these
879  * functions.
880  *
881  * The related inode should be share-locked by the caller.  The caller is
882  * on the frontend.
883  *
884  * record->leaf.base.key
885  * record->leaf.base.rec_type
886  * record->leaf.base.localization
887  */
888 int
889 hammer_ip_add_record(hammer_transaction_t trans, hammer_record_t record)
890 {
891         hammer_inode_t ip = record->ip;
892         int error;
893
894         KKASSERT(record->leaf.base.localization != 0);
895         record->leaf.base.obj_id = ip->obj_id;
896         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
897         error = hammer_mem_add(record);
898         return(error);
899 }
900
901 /*
902  * Locate a pre-existing bulk record in memory.  The caller wishes to
903  * replace the record with a new one.  The existing record may have a
904  * different length (and thus a different key) so we have to use an
905  * overlap check function.
906  */
907 static hammer_record_t
908 hammer_ip_get_bulk(hammer_record_t record)
909 {
910         struct hammer_bulk_info info;
911         hammer_inode_t ip = record->ip;
912
913         info.record = record;
914         info.conflict = NULL;
915         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp,
916                                    hammer_bulk_scan_callback, &info);
917
918         return(info.conflict);  /* may be NULL */
919 }
920
921 /*
922  * Take records vetted by overlap_cmp.  The first non-deleted record
923  * (if any) stops the scan.
924  */
925 static int
926 hammer_bulk_scan_callback(hammer_record_t record, void *data)
927 {
928         struct hammer_bulk_info *info = data;
929
930         if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
931                              HAMMER_RECF_COMMITTED)) {
932                 return(0);
933         }
934         hammer_ref(&record->lock);
935         info->conflict = record;
936         return(-1);                     /* stop scan */
937 }
938
939 /*
940  * Reserve blockmap space placemarked with an in-memory record.
941  *
942  * This routine is called by the frontend in order to be able to directly
943  * flush a buffer cache buffer.  The frontend has locked the related buffer
944  * cache buffers and we should be able to manipulate any overlapping
945  * in-memory records.
946  *
947  * The caller is responsible for adding the returned record and deleting
948  * the returned conflicting record (if any), typically by calling
949  * hammer_ip_replace_bulk() (via hammer_io_direct_write()).
950  */
951 hammer_record_t
952 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
953                    int *errorp)
954 {
955         hammer_record_t record;
956         int zone;
957
958         /*
959          * Create a record to cover the direct write.  The record cannot
960          * be added to the in-memory RB tree here as it might conflict
961          * with an existing memory record.  See hammer_io_direct_write().
962          *
963          * The backend is responsible for finalizing the space reserved in
964          * this record.
965          *
966          * XXX bytes not aligned, depend on the reservation code to
967          * align the reservation.
968          */
969         record = hammer_alloc_mem_record(ip, 0);
970         zone = hammer_data_zone_index(bytes);
971         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
972                                                &record->leaf.data_offset,
973                                                errorp);
974         if (record->resv == NULL) {
975                 hdkprintf("reservation failed\n");
976                 hammer_rel_mem_record(record);
977                 return(NULL);
978         }
979         record->type = HAMMER_MEM_RECORD_DATA;
980         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
981         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
982         record->leaf.base.obj_id = ip->obj_id;
983         record->leaf.base.key = file_offset + bytes;
984         record->leaf.base.localization = ip->obj_localization |
985                                          HAMMER_LOCALIZE_MISC;
986         record->leaf.data_len = bytes;
987         hammer_crc_set_leaf(ip->hmp->version, data, &record->leaf);
988         KKASSERT(*errorp == 0);
989
990         return(record);
991 }
992
993 /*
994  * Called by hammer_io_direct_write() prior to any possible completion
995  * of the BIO to emplace the memory record associated with the I/O and
996  * to replace any prior memory record which might still be active.
997  *
998  * Setting the FE deleted flag on the old record (if any) avoids any RB
999  * tree insertion conflict, amoung other things.
1000  *
1001  * This has to be done prior to the caller completing any related buffer
1002  * cache I/O or a reinstantiation of the buffer may load data from the
1003  * old media location instead of the new media location.  The holding
1004  * of the locked buffer cache buffer serves to interlock the record
1005  * replacement operation.
1006  */
1007 void
1008 hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record)
1009 {
1010         hammer_record_t conflict;
1011         int error __debugvar;
1012
1013         while ((conflict = hammer_ip_get_bulk(record)) != NULL) {
1014                 if ((conflict->flags & HAMMER_RECF_INTERLOCK_BE) == 0) {
1015                         conflict->flags |= HAMMER_RECF_DELETED_FE;
1016                         break;
1017                 }
1018                 conflict->flags |= HAMMER_RECF_WANTED;
1019                 tsleep(conflict, 0, "hmrrc3", 0);
1020                 hammer_rel_mem_record(conflict);
1021         }
1022         error = hammer_mem_add(record);
1023         if (conflict)
1024                 hammer_rel_mem_record(conflict);
1025         KKASSERT(error == 0);
1026 }
1027
1028 /*
1029  * Frontend truncation code.  Scan in-memory records only.  On-disk records
1030  * and records in a flushing state are handled by the backend.  The vnops
1031  * setattr code will handle the block containing the truncation point.
1032  *
1033  * Partial blocks are not deleted.
1034  *
1035  * This code is only called on regular files.
1036  */
1037 int
1038 hammer_ip_frontend_trunc(hammer_inode_t ip, off_t file_size)
1039 {
1040         struct rec_trunc_info info;
1041
1042         switch(ip->ino_data.obj_type) {
1043         case HAMMER_OBJTYPE_REGFILE:
1044                 info.rec_type = HAMMER_RECTYPE_DATA;
1045                 break;
1046         case HAMMER_OBJTYPE_DBFILE:
1047                 info.rec_type = HAMMER_RECTYPE_DB;
1048                 break;
1049         default:
1050                 return(EINVAL);
1051         }
1052         info.trunc_off = file_size;
1053         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
1054                                    hammer_frontend_trunc_callback, &info);
1055         return(0);
1056 }
1057
1058 /*
1059  * Scan callback for frontend records to destroy during a truncation.
1060  * We must ensure that DELETED_FE is set on the record or the frontend
1061  * will get confused in future read() calls.
1062  *
1063  * NOTE: DELETED_FE cannot be set while the record interlock (BE) is held.
1064  *       In this rare case we must wait for the interlock to be cleared.
1065  *
1066  * NOTE: This function is only called on regular files.  There are further
1067  *       restrictions to the setting of DELETED_FE on directory records
1068  *       undergoing a flush due to sensitive inode link count calculations.
1069  */
1070 static int
1071 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
1072 {
1073         if (record->flags & HAMMER_RECF_DELETED_FE)
1074                 return(0);
1075 #if 0
1076         if (record->flush_state == HAMMER_FST_FLUSH)
1077                 return(0);
1078 #endif
1079         hammer_ref(&record->lock);
1080         while (record->flags & HAMMER_RECF_INTERLOCK_BE)
1081                 hammer_wait_mem_record_ident(record, "hmmtrr");
1082         record->flags |= HAMMER_RECF_DELETED_FE;
1083         hammer_rel_mem_record(record);
1084         return(0);
1085 }
1086
1087 /*
1088  * Return 1 if the caller must check for and delete existing records
1089  * before writing out a new data record.
1090  *
1091  * Return 0 if the caller can just insert the record into the B-Tree without
1092  * checking.
1093  */
1094 static int
1095 hammer_record_needs_overwrite_delete(hammer_record_t record)
1096 {
1097         hammer_inode_t ip = record->ip;
1098         int64_t file_offset;
1099         int r;
1100
1101         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
1102                 file_offset = record->leaf.base.key;
1103         else
1104                 file_offset = record->leaf.base.key - record->leaf.data_len;
1105         r = (file_offset < ip->save_trunc_off);
1106         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1107                 if (ip->save_trunc_off <= record->leaf.base.key)
1108                         ip->save_trunc_off = record->leaf.base.key + 1;
1109         } else {
1110                 if (ip->save_trunc_off < record->leaf.base.key)
1111                         ip->save_trunc_off = record->leaf.base.key;
1112         }
1113         return(r);
1114 }
1115
1116 /*
1117  * Backend code.  Sync a record to the media.
1118  */
1119 int
1120 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1121 {
1122         hammer_transaction_t trans = cursor->trans;
1123         hammer_mount_t hmp = trans->hmp;
1124         int64_t file_offset;
1125         int bytes;
1126         void *bdata;
1127         int error;
1128         int doprop;
1129
1130         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1131         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1132         KKASSERT(record->leaf.base.localization != 0);
1133
1134         /*
1135          * Any direct-write related to the record must complete before we
1136          * can sync the record to the on-disk media.
1137          */
1138         if (record->gflags & (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL))
1139                 hammer_io_direct_wait(record);
1140
1141         /*
1142          * If this is a bulk-data record placemarker there may be an existing
1143          * record on-disk, indicating a data overwrite.  If there is the
1144          * on-disk record must be deleted before we can insert our new record.
1145          *
1146          * We've synthesized this record and do not know what the create_tid
1147          * on-disk is, nor how much data it represents.
1148          *
1149          * Keep in mind that (key) for data records is (base_offset + len),
1150          * not (base_offset).  Also, we only want to get rid of on-disk
1151          * records since we are trying to sync our in-memory record, call
1152          * hammer_ip_delete_range() with truncating set to 1 to make sure
1153          * it skips in-memory records.
1154          *
1155          * It is ok for the lookup to return ENOENT.
1156          *
1157          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1158          * to call hammer_ip_delete_range() or not.  This also means we must
1159          * update sync_trunc_off() as we write.
1160          */
1161         if (record->type == HAMMER_MEM_RECORD_DATA &&
1162             hammer_record_needs_overwrite_delete(record)) {
1163                 file_offset = record->leaf.base.key - record->leaf.data_len;
1164                 bytes = HAMMER_BUFSIZE_DOALIGN(record->leaf.data_len);
1165                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1166                 error = hammer_ip_delete_range(
1167                                 cursor, record->ip,
1168                                 file_offset, file_offset + bytes - 1,
1169                                 1);
1170                 if (error && error != ENOENT)
1171                         goto done;
1172         }
1173
1174         /*
1175          * If this is a general record there may be an on-disk version
1176          * that must be deleted before we can insert the new record.
1177          */
1178         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1179                 error = hammer_delete_general(cursor, record->ip, &record->leaf);
1180                 if (error && error != ENOENT)
1181                         goto done;
1182         }
1183
1184         /*
1185          * Setup the cursor.
1186          */
1187         hammer_normalize_cursor(cursor);
1188         cursor->key_beg = record->leaf.base;
1189         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1190         cursor->flags |= HAMMER_CURSOR_BACKEND;
1191         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1192
1193         /*
1194          * Records can wind up on-media before the inode itself is on-media.
1195          * Flag the case.
1196          */
1197         record->ip->flags |= HAMMER_INODE_DONDISK;
1198
1199         /*
1200          * If we are deleting a directory entry an exact match must be
1201          * found on-disk.
1202          */
1203         if (record->type == HAMMER_MEM_RECORD_DEL) {
1204                 error = hammer_btree_lookup(cursor);
1205                 if (error == 0) {
1206                         KKASSERT(cursor->iprec == NULL);
1207                         error = hammer_ip_delete_record(cursor, record->ip,
1208                                                         trans->tid);
1209                         if (error == 0) {
1210                                 record->flags |= HAMMER_RECF_DELETED_BE |
1211                                                  HAMMER_RECF_COMMITTED;
1212                                 ++record->ip->rec_generation;
1213                         }
1214                 }
1215                 goto done;
1216         }
1217
1218         /*
1219          * We are inserting.
1220          *
1221          * Issue a lookup to position the cursor and locate the insertion
1222          * point.  The target key should not exist.  If we are creating a
1223          * directory entry we may have to iterate the low 32 bits of the
1224          * key to find an unused key.
1225          */
1226         hammer_sync_lock_sh(trans);
1227         cursor->flags |= HAMMER_CURSOR_INSERT;
1228         error = hammer_btree_lookup(cursor);
1229         if (hammer_debug_inode)
1230                 hdkprintf("DOINSERT LOOKUP %d\n", error);
1231         if (error == 0) {
1232                 hdkprintf("duplicate rec at (%016jx)\n",
1233                         (intmax_t)record->leaf.base.key);
1234                 if (hammer_debug_critical)
1235                         Debugger("duplicate record1");
1236                 error = EIO;
1237         }
1238
1239         if (error != ENOENT)
1240                 goto done_unlock;
1241
1242         /*
1243          * Allocate the record and data.  The result buffers will be
1244          * marked as being modified and further calls to
1245          * hammer_modify_buffer() will result in unneeded UNDO records.
1246          *
1247          * Support zero-fill records (data == NULL and data_len != 0)
1248          */
1249         if (record->type == HAMMER_MEM_RECORD_DATA) {
1250                 /*
1251                  * The data portion of a bulk-data record has already been
1252                  * committed to disk, we need only adjust the layer2
1253                  * statistics in the same transaction as our B-Tree insert.
1254                  */
1255                 KKASSERT(record->leaf.data_offset != 0);
1256                 error = hammer_blockmap_finalize(trans,
1257                                                  record->resv,
1258                                                  record->leaf.data_offset,
1259                                                  record->leaf.data_len);
1260         } else if (record->data && record->leaf.data_len) {
1261                 /*
1262                  * Wholely cached record, with data.  Allocate the data.
1263                  */
1264                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1265                                           record->leaf.base.rec_type,
1266                                           &record->leaf.data_offset,
1267                                           &cursor->data_buffer,
1268                                           0, &error);
1269                 if (bdata == NULL)
1270                         goto done_unlock;
1271                 hammer_crc_set_leaf(hmp->version, record->data, &record->leaf);
1272                 hammer_modify_buffer_noundo(trans, cursor->data_buffer);
1273                 bcopy(record->data, bdata, record->leaf.data_len);
1274                 hammer_modify_buffer_done(cursor->data_buffer);
1275         } else {
1276                 /*
1277                  * Wholely cached record, without data.
1278                  */
1279                 record->leaf.data_offset = 0;
1280                 record->leaf.data_crc = 0;
1281         }
1282
1283         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1284         if (hammer_debug_inode && error) {
1285                 hdkprintf("BTREE INSERT error %d @ %016jx:%d key %016jx\n",
1286                         error,
1287                         (intmax_t)cursor->node->node_offset,
1288                         cursor->index,
1289                         (intmax_t)record->leaf.base.key);
1290         }
1291
1292         /*
1293          * Our record is on-disk and we normally mark the in-memory version
1294          * as having been committed (and not BE-deleted).
1295          *
1296          * If the record represented a directory deletion but we had to
1297          * sync a valid directory entry to disk due to dependancies,
1298          * we must convert the record to a covering delete so the
1299          * frontend does not have visibility on the synced entry.
1300          *
1301          * WARNING: cursor's leaf pointer may have changed after do_propagation
1302          *          returns!
1303          */
1304         if (error == 0) {
1305                 if (doprop) {
1306                         hammer_btree_do_propagation(cursor, &record->leaf);
1307                 }
1308                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1309                         /*
1310                          * Must convert deleted directory entry add
1311                          * to a directory entry delete.
1312                          */
1313                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1314                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1315                         record->type = HAMMER_MEM_RECORD_DEL;
1316                         KKASSERT(record->ip->obj_id == record->leaf.base.obj_id);
1317                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1318                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1319                         KKASSERT((record->flags & (HAMMER_RECF_COMMITTED |
1320                                                  HAMMER_RECF_DELETED_BE)) == 0);
1321                         /* converted record is not yet committed */
1322                         /* hammer_flush_record_done takes care of the rest */
1323                 } else {
1324                         /*
1325                          * Everything went fine and we are now done with
1326                          * this record.
1327                          */
1328                         record->flags |= HAMMER_RECF_COMMITTED;
1329                         ++record->ip->rec_generation;
1330                 }
1331         } else {
1332                 if (record->leaf.data_offset) {
1333                         hammer_blockmap_free(trans, record->leaf.data_offset,
1334                                              record->leaf.data_len);
1335                 }
1336         }
1337 done_unlock:
1338         hammer_sync_unlock(trans);
1339 done:
1340         return(error);
1341 }
1342
1343 /*
1344  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1345  * entry's key is used to deal with hash collisions in the upper 32 bits.
1346  * A unique 64 bit key is generated in-memory and may be regenerated a
1347  * second time when the directory record is flushed to the on-disk B-Tree.
1348  *
1349  * A referenced record is passed to this function.  This function
1350  * eats the reference.  If an error occurs the record will be deleted.
1351  *
1352  * A copy of the temporary record->data pointer provided by the caller
1353  * will be made.
1354  */
1355 int
1356 hammer_mem_add(hammer_record_t record)
1357 {
1358         hammer_mount_t hmp = record->ip->hmp;
1359
1360         /*
1361          * Make a private copy of record->data
1362          */
1363         if (record->data)
1364                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1365
1366         /*
1367          * Insert into the RB tree.  A unique key should have already
1368          * been selected if this is a directory entry.
1369          */
1370         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1371                 record->flags |= HAMMER_RECF_DELETED_FE;
1372                 hammer_rel_mem_record(record);
1373                 return (EEXIST);
1374         }
1375         ++hmp->rsv_recs;
1376         ++record->ip->rsv_recs;
1377         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1378         record->flags |= HAMMER_RECF_ONRBTREE;
1379         hammer_modify_inode(NULL, record->ip, HAMMER_INODE_XDIRTY);
1380         hammer_rel_mem_record(record);
1381         return(0);
1382 }
1383
1384 /************************************************************************
1385  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1386  ************************************************************************
1387  *
1388  * These functions augment the B-Tree scanning functions in hammer_btree.c
1389  * by merging in-memory records with on-disk records.
1390  */
1391
1392 /*
1393  * Locate a particular record either in-memory or on-disk.
1394  *
1395  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1396  * NOT be called to iterate results.
1397  */
1398 int
1399 hammer_ip_lookup(hammer_cursor_t cursor)
1400 {
1401         int error;
1402
1403         /*
1404          * If the element is in-memory return it without searching the
1405          * on-disk B-Tree
1406          */
1407         KKASSERT(cursor->ip);
1408         error = hammer_mem_lookup(cursor);
1409         if (error == 0) {
1410                 cursor->leaf = &cursor->iprec->leaf;
1411                 return(error);
1412         }
1413         if (error != ENOENT)
1414                 return(error);
1415
1416         /*
1417          * If the inode has on-disk components search the on-disk B-Tree.
1418          */
1419         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1420                 return(error);
1421         error = hammer_btree_lookup(cursor);
1422         if (error == 0)
1423                 error = hammer_btree_extract_leaf(cursor);
1424         return(error);
1425 }
1426
1427 /*
1428  * Helper for hammer_ip_first()/hammer_ip_next()
1429  *
1430  * NOTE: Both ATEDISK and DISKEOF will be set the same.  This sets up
1431  * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek
1432  * state if hammer_ip_next() needs to re-seek.
1433  */
1434 static __inline
1435 int
1436 _hammer_ip_seek_btree(hammer_cursor_t cursor)
1437 {
1438         hammer_inode_t ip = cursor->ip;
1439         int error;
1440
1441         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1442                 error = hammer_btree_lookup(cursor);
1443                 if (error == ENOENT || error == EDEADLK) {
1444                         if (hammer_debug_general & 0x2000) {
1445                                 hdkprintf("error %d node %p %016jx index %d\n",
1446                                         error, cursor->node,
1447                                         (intmax_t)cursor->node->node_offset,
1448                                         cursor->index);
1449                         }
1450                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1451                         error = hammer_btree_iterate(cursor);
1452                 }
1453                 if (error == 0) {
1454                         cursor->flags &= ~(HAMMER_CURSOR_DISKEOF |
1455                                            HAMMER_CURSOR_ATEDISK);
1456                 } else {
1457                         cursor->flags |= HAMMER_CURSOR_DISKEOF |
1458                                          HAMMER_CURSOR_ATEDISK;
1459                         if (error == ENOENT)
1460                                 error = 0;
1461                 }
1462         } else {
1463                 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK;
1464                 error = 0;
1465         }
1466         return(error);
1467 }
1468
1469 /*
1470  * Helper for hammer_ip_next()
1471  *
1472  * The caller has determined that the media cursor is further along than the
1473  * memory cursor and must be reseeked after a generation number change.
1474  */
1475 static
1476 int
1477 _hammer_ip_reseek(hammer_cursor_t cursor)
1478 {
1479         struct hammer_base_elm save;
1480         hammer_btree_elm_t elm;
1481         int error __debugvar;
1482         int r;
1483         int again = 0;
1484
1485         /*
1486          * Do the re-seek.
1487          */
1488         hkprintf("Debug: re-seeked during scan @ino=%016jx\n",
1489                 (intmax_t)cursor->ip->obj_id);
1490         save = cursor->key_beg;
1491         cursor->key_beg = cursor->iprec->leaf.base;
1492         error = _hammer_ip_seek_btree(cursor);
1493         KKASSERT(error == 0);
1494         cursor->key_beg = save;
1495
1496         /*
1497          * If the memory record was previous returned to
1498          * the caller and the media record matches
1499          * (-1/+1: only create_tid differs), then iterate
1500          * the media record to avoid a double result.
1501          */
1502         if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 &&
1503             (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) {
1504                 elm = &cursor->node->ondisk->elms[cursor->index];
1505                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1506                 if (cursor->flags & HAMMER_CURSOR_ASOF) {
1507                         if (r >= -1 && r <= 1) {
1508                                 hkprintf("Debug: iterated after "
1509                                         "re-seek (asof r=%d)\n", r);
1510                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1511                                 again = 1;
1512                         }
1513                 } else {
1514                         if (r == 0) {
1515                                 hkprintf("Debug: iterated after "
1516                                         "re-seek\n");
1517                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1518                                 again = 1;
1519                         }
1520                 }
1521         }
1522         return(again);
1523 }
1524
1525 /*
1526  * Locate the first record within the cursor's key_beg/key_end range,
1527  * restricted to a particular inode.  0 is returned on success, ENOENT
1528  * if no records matched the requested range, or some other error.
1529  *
1530  * When 0 is returned hammer_ip_next() may be used to iterate additional
1531  * records within the requested range.
1532  *
1533  * This function can return EDEADLK, requiring the caller to terminate
1534  * the cursor and try again.
1535  */
1536
1537 int
1538 hammer_ip_first(hammer_cursor_t cursor)
1539 {
1540         hammer_inode_t ip __debugvar = cursor->ip;
1541         int error;
1542
1543         KKASSERT(ip != NULL);
1544
1545         /*
1546          * Clean up fields and setup for merged scan
1547          */
1548         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1549
1550         /*
1551          * Search the in-memory record list (Red-Black tree).  Unlike the
1552          * B-Tree search, mem_first checks for records in the range.
1553          *
1554          * This function will setup both ATEMEM and MEMEOF properly for
1555          * the ip iteration.  ATEMEM will be set if MEMEOF is set.
1556          */
1557         hammer_mem_first(cursor);
1558
1559         /*
1560          * Detect generation changes during blockages, including
1561          * blockages which occur on the initial btree search.
1562          */
1563         cursor->rec_generation = cursor->ip->rec_generation;
1564
1565         /*
1566          * Initial search and result
1567          */
1568         error = _hammer_ip_seek_btree(cursor);
1569         if (error == 0)
1570                 error = hammer_ip_next(cursor);
1571
1572         return (error);
1573 }
1574
1575 /*
1576  * Retrieve the next record in a merged iteration within the bounds of the
1577  * cursor.  This call may be made multiple times after the cursor has been
1578  * initially searched with hammer_ip_first().
1579  *
1580  * There are numerous special cases in this code to deal with races between
1581  * in-memory records and on-media records.
1582  *
1583  * 0 is returned on success, ENOENT if no further records match the
1584  * requested range, or some other error code is returned.
1585  */
1586 int
1587 hammer_ip_next(hammer_cursor_t cursor)
1588 {
1589         hammer_btree_elm_t elm;
1590         hammer_record_t rec;
1591         hammer_record_t tmprec;
1592         int error;
1593         int r;
1594
1595 again:
1596         /*
1597          * Get the next on-disk record
1598          *
1599          * NOTE: If we deleted the last on-disk record we had scanned
1600          *       ATEDISK will be clear and RETEST will be set, forcing
1601          *       a call to iterate.  The fact that ATEDISK is clear causes
1602          *       iterate to re-test the 'current' element.  If ATEDISK is
1603          *       set, iterate will skip the 'current' element.
1604          */
1605         error = 0;
1606         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1607                 if (cursor->flags & (HAMMER_CURSOR_ATEDISK |
1608                                      HAMMER_CURSOR_RETEST)) {
1609                         error = hammer_btree_iterate(cursor);
1610                         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1611                         if (error == 0) {
1612                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1613                                 hammer_cache_node(&cursor->ip->cache[1],
1614                                                   cursor->node);
1615                         } else if (error == ENOENT) {
1616                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1617                                                  HAMMER_CURSOR_ATEDISK;
1618                                 error = 0;
1619                         }
1620                 }
1621         }
1622
1623         /*
1624          * If the generation changed the backend has deleted or committed
1625          * one or more memory records since our last check.
1626          *
1627          * When this case occurs if the disk cursor is > current memory record
1628          * or the disk cursor is at EOF, we must re-seek the disk-cursor.
1629          * Since the cursor is ahead it must have not yet been eaten (if
1630          * not at eof anyway). (XXX data offset case?)
1631          *
1632          * NOTE: we are not doing a full check here.  That will be handled
1633          * later on.
1634          *
1635          * If we have exhausted all memory records we do not have to do any
1636          * further seeks.
1637          */
1638         while (cursor->rec_generation != cursor->ip->rec_generation &&
1639                error == 0) {
1640                 hkprintf("Debug: generation changed during scan @ino=%016jx\n",
1641                         (intmax_t)cursor->ip->obj_id);
1642                 cursor->rec_generation = cursor->ip->rec_generation;
1643                 if (cursor->flags & HAMMER_CURSOR_MEMEOF)
1644                         break;
1645                 if (cursor->flags & HAMMER_CURSOR_DISKEOF) {
1646                         r = 1;
1647                 } else {
1648                         KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0);
1649                         elm = &cursor->node->ondisk->elms[cursor->index];
1650                         r = hammer_btree_cmp(&elm->base,
1651                                              &cursor->iprec->leaf.base);
1652                 }
1653
1654                 /*
1655                  * Do we re-seek the media cursor?
1656                  */
1657                 if (r > 0) {
1658                         if (_hammer_ip_reseek(cursor))
1659                                 goto again;
1660                 }
1661         }
1662
1663         /*
1664          * We can now safely get the next in-memory record.  We cannot
1665          * block here.
1666          *
1667          * hammer_rec_scan_cmp:  Is the record still in our general range,
1668          *                       (non-inclusive of snapshot exclusions)?
1669          * hammer_rec_scan_callback: Is the record in our snapshot?
1670          */
1671         tmprec = NULL;
1672         if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1673                 /*
1674                  * If the current memory record was eaten then get the next
1675                  * one.  Stale records are skipped.
1676                  */
1677                 if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1678                         tmprec = cursor->iprec;
1679                         cursor->iprec = NULL;
1680                         rec = hammer_rec_rb_tree_RB_NEXT(tmprec);
1681                         while (rec) {
1682                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1683                                         break;
1684                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1685                                         break;
1686                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1687                         }
1688                         if (cursor->iprec) {
1689                                 KKASSERT(cursor->iprec == rec);
1690                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1691                         } else {
1692                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1693                         }
1694                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1695                 }
1696         }
1697
1698         /*
1699          * MEMORY RECORD VALIDITY TEST
1700          *
1701          * (We still can't block, which is why tmprec is being held so
1702          * long).
1703          *
1704          * If the memory record is no longer valid we skip it.  It may
1705          * have been deleted by the frontend.  If it was deleted or
1706          * committed by the backend the generation change re-seeked the
1707          * disk cursor and the record will be present there.
1708          */
1709         if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1710                 KKASSERT(cursor->iprec);
1711                 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0);
1712                 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) {
1713                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1714                         if (tmprec)
1715                                 hammer_rel_mem_record(tmprec);
1716                         goto again;
1717                 }
1718         }
1719         if (tmprec)
1720                 hammer_rel_mem_record(tmprec);
1721
1722         /*
1723          * Extract either the disk or memory record depending on their
1724          * relative position.
1725          */
1726         error = 0;
1727         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1728         case 0:
1729                 /*
1730                  * Both entries valid.   Compare the entries and nominally
1731                  * return the first one in the sort order.  Numerous cases
1732                  * require special attention, however.
1733                  */
1734                 elm = &cursor->node->ondisk->elms[cursor->index];
1735                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1736
1737                 /*
1738                  * If the two entries differ only by their key (-2/2) or
1739                  * create_tid (-1/1), and are DATA records, we may have a
1740                  * nominal match.  We have to calculate the base file
1741                  * offset of the data.
1742                  */
1743                 if (r <= 2 && r >= -2 && r != 0 &&
1744                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1745                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1746                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1747                         int64_t base2 = cursor->iprec->leaf.base.key -
1748                                         cursor->iprec->leaf.data_len;
1749                         if (base1 == base2)
1750                                 r = 0;
1751                 }
1752
1753                 if (r < 0) {
1754                         error = hammer_btree_extract_leaf(cursor);
1755                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1756                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1757                         break;
1758                 }
1759
1760                 /*
1761                  * If the entries match exactly the memory entry is either
1762                  * an on-disk directory entry deletion or a bulk data
1763                  * overwrite.  If it is a directory entry deletion we eat
1764                  * both entries.
1765                  *
1766                  * For the bulk-data overwrite case it is possible to have
1767                  * visibility into both, which simply means the syncer
1768                  * hasn't gotten around to doing the delete+insert sequence
1769                  * on the B-Tree.  Use the memory entry and throw away the
1770                  * on-disk entry.
1771                  *
1772                  * If the in-memory record is not either of these we
1773                  * probably caught the syncer while it was syncing it to
1774                  * the media.  Since we hold a shared lock on the cursor,
1775                  * the in-memory record had better be marked deleted at
1776                  * this point.
1777                  */
1778                 if (r == 0) {
1779                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1780                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1781                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1782                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1783                                         goto again;
1784                                 }
1785                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1786                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1787                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1788                                 }
1789                                 /* fall through to memory entry */
1790                         } else {
1791                                 hpanic("duplicate mem/B-Tree entry %p %d %08x",
1792                                         cursor->iprec,
1793                                         cursor->iprec->type,
1794                                         cursor->iprec->flags);
1795                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1796                                 goto again;
1797                         }
1798                 }
1799                 /* fall through to the memory entry */
1800         case HAMMER_CURSOR_ATEDISK:
1801                 /*
1802                  * Only the memory entry is valid.
1803                  */
1804                 cursor->leaf = &cursor->iprec->leaf;
1805                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1806                 cursor->flags |= HAMMER_CURSOR_LASTWASMEM;
1807
1808                 /*
1809                  * If the memory entry is an on-disk deletion we should have
1810                  * also had found a B-Tree record.  If the backend beat us
1811                  * to it it would have interlocked the cursor and we should
1812                  * have seen the in-memory record marked DELETED_FE.
1813                  */
1814                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1815                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1816                         hpanic("del-on-disk with no B-Tree entry iprec %p flags %08x",
1817                                 cursor->iprec,
1818                                 cursor->iprec->flags);
1819                 }
1820                 break;
1821         case HAMMER_CURSOR_ATEMEM:
1822                 /*
1823                  * Only the disk entry is valid
1824                  */
1825                 error = hammer_btree_extract_leaf(cursor);
1826                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1827                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1828                 break;
1829         default:
1830                 /*
1831                  * Neither entry is valid
1832                  *
1833                  * XXX error not set properly
1834                  */
1835                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1836                 cursor->leaf = NULL;
1837                 error = ENOENT;
1838                 break;
1839         }
1840         return(error);
1841 }
1842
1843 /*
1844  * Resolve the cursor->data pointer for the current cursor position in
1845  * a merged iteration.
1846  */
1847 int
1848 hammer_ip_resolve_data(hammer_cursor_t cursor)
1849 {
1850         hammer_record_t record;
1851         int error;
1852
1853         if (hammer_cursor_inmem(cursor)) {
1854                 /*
1855                  * The data associated with an in-memory record is usually
1856                  * kmalloced, but reserve-ahead data records will have an
1857                  * on-disk reference.
1858                  *
1859                  * NOTE: Reserve-ahead data records must be handled in the
1860                  *       context of the related high level buffer cache buffer
1861                  *       to interlock against async writes.
1862                  *
1863                  * NOTE: We might catch a direct write in-progress, in which
1864                  *       case we must wait for it to complete.  The wait
1865                  *       function will also clean out any buffer aliases.
1866                  *
1867                  *       (In fact, it is possible that the write had not
1868                  *        even started yet).
1869                  */
1870                 record = cursor->iprec;
1871                 cursor->data = record->data;
1872                 error = 0;
1873                 if (cursor->data == NULL) {
1874                         hammer_io_direct_wait(record);
1875                         KKASSERT(record->leaf.base.rec_type ==
1876                                  HAMMER_RECTYPE_DATA);
1877                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1878                                                     record->leaf.data_offset,
1879                                                     record->leaf.data_len,
1880                                                     &error,
1881                                                     &cursor->data_buffer);
1882                 }
1883         } else {
1884                 /*
1885                  * Loading leaf here isn't necessary if it's guaranteed that
1886                  * the cursor is at a leaf node (which basically should be)
1887                  * because hammer_btree_extract_data() does that.
1888                  */
1889                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1890                 error = hammer_btree_extract_data(cursor);
1891         }
1892         return(error);
1893 }
1894
1895 /*
1896  * Backend truncation / record replacement - delete records in range.
1897  *
1898  * Delete all records within the specified range for inode ip.  In-memory
1899  * records still associated with the frontend are ignored.
1900  *
1901  * If truncating is non-zero in-memory records associated with the back-end
1902  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1903  *
1904  * NOTES:
1905  *
1906  *      * An unaligned range will cause new records to be added to cover
1907  *        the edge cases. (XXX not implemented yet).
1908  *
1909  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1910  *        also do not deal with unaligned ranges.
1911  *
1912  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1913  *
1914  *      * Record keys for regular file data have to be special-cased since
1915  *        they indicate the end of the range (key = base + bytes).
1916  *
1917  *      * This function may be asked to delete ridiculously huge ranges, for
1918  *        example if someone truncates or removes a 1TB regular file.  We
1919  *        must be very careful on restarts and we may have to stop w/
1920  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1921  */
1922 int
1923 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1924                        int64_t ran_beg, int64_t ran_end, int truncating)
1925 {
1926         hammer_transaction_t trans = cursor->trans;
1927         hammer_btree_leaf_elm_t leaf;
1928         int error;
1929         int64_t off;
1930         int64_t tmp64;
1931
1932         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1933 retry:
1934         hammer_normalize_cursor(cursor);
1935         cursor->key_beg.localization = ip->obj_localization |
1936                                        HAMMER_LOCALIZE_MISC;
1937         cursor->key_beg.obj_id = ip->obj_id;
1938         cursor->key_beg.create_tid = 0;
1939         cursor->key_beg.delete_tid = 0;
1940         cursor->key_beg.obj_type = 0;
1941
1942         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1943                 cursor->key_beg.key = ran_beg;
1944                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1945         } else {
1946                 /*
1947                  * The key in the B-Tree is (base+bytes), so the first possible
1948                  * matching key is ran_beg + 1.
1949                  */
1950                 cursor->key_beg.key = ran_beg + 1;
1951                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1952         }
1953
1954         cursor->key_end = cursor->key_beg;
1955         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1956                 cursor->key_end.key = ran_end;
1957         } else {
1958                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1959                 if (tmp64 < ran_end)
1960                         cursor->key_end.key = HAMMER_MAX_KEY;
1961                 else
1962                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1963         }
1964
1965         cursor->asof = ip->obj_asof;
1966         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1967         cursor->flags |= HAMMER_CURSOR_ASOF;
1968         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1969         cursor->flags |= HAMMER_CURSOR_BACKEND;
1970         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1971
1972         error = hammer_ip_first(cursor);
1973
1974         /*
1975          * Iterate through matching records and mark them as deleted.
1976          */
1977         while (error == 0) {
1978                 leaf = cursor->leaf;
1979
1980                 KKASSERT(leaf->base.delete_tid == 0);
1981                 KKASSERT(leaf->base.obj_id == ip->obj_id);
1982
1983                 /*
1984                  * There may be overlap cases for regular file data.  Also
1985                  * remember the key for a regular file record is (base + len),
1986                  * NOT (base).
1987                  *
1988                  * Note that due to duplicates (mem & media) allowed by
1989                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
1990                  */
1991                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1992                         off = leaf->base.key - leaf->data_len;
1993                         /*
1994                          * Check the left edge case.  We currently do not
1995                          * split existing records.
1996                          */
1997                         if (off < ran_beg && leaf->base.key > ran_beg) {
1998                                 hpanic("hammer left edge case %016jx %d",
1999                                         (intmax_t)leaf->base.key,
2000                                         leaf->data_len);
2001                         }
2002
2003                         /*
2004                          * Check the right edge case.  Note that the
2005                          * record can be completely out of bounds, which
2006                          * terminates the search.
2007                          *
2008                          * base->key is exclusive of the right edge while
2009                          * ran_end is inclusive of the right edge.  The
2010                          * (key - data_len) left boundary is inclusive.
2011                          *
2012                          * XXX theory-check this test at some point, are
2013                          * we missing a + 1 somewhere?  Note that ran_end
2014                          * could overflow.
2015                          */
2016                         if (leaf->base.key - 1 > ran_end) {
2017                                 if (leaf->base.key - leaf->data_len > ran_end)
2018                                         break;
2019                                 hpanic("hammer right edge case");
2020                         }
2021                 } else {
2022                         off = leaf->base.key;
2023                 }
2024
2025                 /*
2026                  * Delete the record.  When truncating we do not delete
2027                  * in-memory (data) records because they represent data
2028                  * written after the truncation.
2029                  *
2030                  * This will also physically destroy the B-Tree entry and
2031                  * data if the retention policy dictates.  The function
2032                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2033                  * to retest the new 'current' element.
2034                  */
2035                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
2036                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
2037                         /*
2038                          * If we have built up too many meta-buffers we risk
2039                          * deadlocking the kernel and must stop.  This can
2040                          * occur when deleting ridiculously huge files.
2041                          * sync_trunc_off is updated so the next cycle does
2042                          * not re-iterate records we have already deleted.
2043                          *
2044                          * This is only done with formal truncations.
2045                          */
2046                         if (truncating > 1 && error == 0 &&
2047                             hammer_flusher_meta_limit(ip->hmp)) {
2048                                 ip->sync_trunc_off = off;
2049                                 error = EWOULDBLOCK;
2050                         }
2051                 }
2052                 if (error)
2053                         break;
2054                 ran_beg = off;  /* for restart */
2055                 error = hammer_ip_next(cursor);
2056         }
2057         if (cursor->node)
2058                 hammer_cache_node(&ip->cache[1], cursor->node);
2059
2060         if (error == EDEADLK) {
2061                 hammer_done_cursor(cursor);
2062                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2063                 if (error == 0)
2064                         goto retry;
2065         }
2066         if (error == ENOENT)
2067                 error = 0;
2068         return(error);
2069 }
2070
2071 /*
2072  * This backend function deletes the specified record on-disk, similar to
2073  * delete_range but for a specific record.  Unlike the exact deletions
2074  * used when deleting a directory entry this function uses an ASOF search
2075  * like delete_range.
2076  *
2077  * This function may be called with ip->obj_asof set for a slave snapshot,
2078  * so don't use it.  We always delete non-historical records only.
2079  */
2080 static int
2081 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
2082                       hammer_btree_leaf_elm_t leaf)
2083 {
2084         hammer_transaction_t trans = cursor->trans;
2085         int error;
2086
2087         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2088 retry:
2089         hammer_normalize_cursor(cursor);
2090         cursor->key_beg = leaf->base;
2091         cursor->asof = HAMMER_MAX_TID;
2092         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2093         cursor->flags |= HAMMER_CURSOR_ASOF;
2094         cursor->flags |= HAMMER_CURSOR_BACKEND;
2095         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2096
2097         error = hammer_btree_lookup(cursor);
2098         if (error == 0) {
2099                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2100         }
2101         if (error == EDEADLK) {
2102                 hammer_done_cursor(cursor);
2103                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2104                 if (error == 0)
2105                         goto retry;
2106         }
2107         return(error);
2108 }
2109
2110 /*
2111  * This function deletes remaining auxillary records when an inode is
2112  * being deleted.  This function explicitly does not delete the
2113  * inode record, directory entry, data, or db records.  Those must be
2114  * properly disposed of prior to this call.
2115  */
2116 int
2117 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
2118 {
2119         hammer_transaction_t trans = cursor->trans;
2120         hammer_btree_leaf_elm_t leaf __debugvar;
2121         int error;
2122
2123         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2124 retry:
2125         hammer_normalize_cursor(cursor);
2126         cursor->key_beg.localization = ip->obj_localization |
2127                                        HAMMER_LOCALIZE_MISC;
2128         cursor->key_beg.obj_id = ip->obj_id;
2129         cursor->key_beg.create_tid = 0;
2130         cursor->key_beg.delete_tid = 0;
2131         cursor->key_beg.obj_type = 0;
2132         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
2133         cursor->key_beg.key = HAMMER_MIN_KEY;
2134
2135         cursor->key_end = cursor->key_beg;
2136         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
2137         cursor->key_end.key = HAMMER_MAX_KEY;
2138
2139         cursor->asof = ip->obj_asof;
2140         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2141         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2142         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2143         cursor->flags |= HAMMER_CURSOR_BACKEND;
2144
2145         error = hammer_ip_first(cursor);
2146
2147         /*
2148          * Iterate through matching records and mark them as deleted.
2149          */
2150         while (error == 0) {
2151                 leaf = cursor->leaf;
2152
2153                 KKASSERT(leaf->base.delete_tid == 0);
2154
2155                 /*
2156                  * Mark the record and B-Tree entry as deleted.  This will
2157                  * also physically delete the B-Tree entry, record, and
2158                  * data if the retention policy dictates.  The function
2159                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2160                  * to retest the new 'current' element.
2161                  *
2162                  * Directory entries (and delete-on-disk directory entries)
2163                  * must be synced and cannot be deleted.
2164                  */
2165                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2166                 ++*countp;
2167                 if (error)
2168                         break;
2169                 error = hammer_ip_next(cursor);
2170         }
2171         if (cursor->node)
2172                 hammer_cache_node(&ip->cache[1], cursor->node);
2173         if (error == EDEADLK) {
2174                 hammer_done_cursor(cursor);
2175                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2176                 if (error == 0)
2177                         goto retry;
2178         }
2179         if (error == ENOENT)
2180                 error = 0;
2181         return(error);
2182 }
2183
2184 /*
2185  * Delete the record at the current cursor.  On success the cursor will
2186  * be positioned appropriately for an iteration but may no longer be at
2187  * a leaf node.
2188  *
2189  * This routine is only called from the backend.
2190  *
2191  * NOTE: This can return EDEADLK, requiring the caller to terminate the
2192  * cursor and retry.
2193  */
2194 int
2195 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
2196                         hammer_tid_t tid)
2197 {
2198         hammer_record_t iprec;
2199         int error;
2200
2201         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
2202         KKASSERT(tid != 0);
2203
2204         /*
2205          * In-memory (unsynchronized) records can simply be freed.  This
2206          * only occurs in range iterations since all other records are
2207          * individually synchronized.  Thus there should be no confusion with
2208          * the interlock.
2209          *
2210          * An in-memory record may be deleted before being committed to disk,
2211          * but could have been accessed in the mean time.  The reservation
2212          * code will deal with the case.
2213          */
2214         if (hammer_cursor_inmem(cursor)) {
2215                 iprec = cursor->iprec;
2216                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
2217                 iprec->flags |= HAMMER_RECF_DELETED_FE;
2218                 iprec->flags |= HAMMER_RECF_DELETED_BE;
2219                 KKASSERT(iprec->ip == ip);
2220                 ++ip->rec_generation;
2221                 return(0);
2222         }
2223
2224         /*
2225          * On-disk records are marked as deleted by updating their delete_tid.
2226          * This does not effect their position in the B-Tree (which is based
2227          * on their create_tid).
2228          *
2229          * Frontend B-Tree operations track inodes so we tell
2230          * hammer_delete_at_cursor() not to.
2231          */
2232         error = hammer_btree_extract_leaf(cursor);
2233         if (error == 0) {
2234                 error = hammer_delete_at_cursor(
2235                                 cursor,
2236                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
2237                                 cursor->trans->tid,
2238                                 cursor->trans->time32,
2239                                 0, NULL);
2240         }
2241         return(error);
2242 }
2243
2244 /*
2245  * Used to write a generic record w/optional data to the media b-tree
2246  * when no inode context is available.  Used by the mirroring and
2247  * snapshot code.
2248  *
2249  * Caller must set cursor->key_beg to leaf->base.  The cursor must be
2250  * flagged for backend operation and not flagged ASOF (since we are
2251  * doing an insertion).
2252  *
2253  * This function will acquire the appropriate sync lock and will set
2254  * the cursor insertion flag for the operation, do the btree lookup,
2255  * and the insertion, and clear the insertion flag and sync lock before
2256  * returning.  The cursor state will be such that the caller can continue
2257  * scanning (used by the mirroring code).
2258  *
2259  * mode: HAMMER_CREATE_MODE_UMIRROR     copyin data, check crc
2260  *       HAMMER_CREATE_MODE_SYS         bcopy data, generate crc
2261  *
2262  * NOTE: EDEADLK can be returned.  The caller must do deadlock handling and
2263  *                retry.
2264  *
2265  *       EALREADY can be returned if the record already exists (WARNING,
2266  *                because ASOF cannot be used no check is made for illegal
2267  *                duplicates).
2268  *
2269  * NOTE: Do not use the function for normal inode-related records as this
2270  *       functions goes directly to the media and is not integrated with
2271  *       in-memory records.
2272  */
2273 int
2274 hammer_create_at_cursor(hammer_cursor_t cursor, hammer_btree_leaf_elm_t leaf,
2275                         void *udata, int mode)
2276 {
2277         hammer_transaction_t trans;
2278         hammer_mount_t hmp;
2279         hammer_buffer_t data_buffer;
2280         hammer_off_t ndata_offset;
2281         hammer_tid_t high_tid;
2282         void *ndata;
2283         int error;
2284         int doprop;
2285
2286         trans = cursor->trans;
2287         hmp = trans->hmp;
2288         data_buffer = NULL;
2289         ndata_offset = 0;
2290         doprop = 0;
2291
2292         KKASSERT((cursor->flags &
2293                   (HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF)) ==
2294                   (HAMMER_CURSOR_BACKEND));
2295
2296         hammer_sync_lock_sh(trans);
2297
2298         if (leaf->data_len) {
2299                 ndata = hammer_alloc_data(trans, leaf->data_len,
2300                                           leaf->base.rec_type,
2301                                           &ndata_offset, &data_buffer,
2302                                           0, &error);
2303                 if (ndata == NULL) {
2304                         hammer_sync_unlock(trans);
2305                         return (error);
2306                 }
2307                 leaf->data_offset = ndata_offset;
2308                 hammer_modify_buffer_noundo(trans, data_buffer);
2309
2310                 switch(mode) {
2311                 case HAMMER_CREATE_MODE_UMIRROR:
2312                         error = copyin(udata, ndata, leaf->data_len);
2313                         if (error == 0) {
2314                                 if (hammer_crc_test_leaf(hmp->version, ndata, leaf) == 0) {
2315                                         hdkprintf("CRC DATA @ %016jx/%d MISMATCH ON PIPE\n",
2316                                                 (intmax_t)ndata_offset,
2317                                                 leaf->data_len);
2318                                         error = EINVAL;
2319                                 } else {
2320                                         error = hammer_cursor_localize_data(
2321                                                         hmp, ndata, leaf);
2322                                 }
2323                         }
2324                         break;
2325                 case HAMMER_CREATE_MODE_SYS:
2326                         bcopy(udata, ndata, leaf->data_len);
2327                         error = 0;
2328                         hammer_crc_set_leaf(hmp->version, ndata, leaf);
2329                         break;
2330                 default:
2331                         hpanic("bad mode %d", mode);
2332                         break; /* NOT REACHED */
2333                 }
2334                 hammer_modify_buffer_done(data_buffer);
2335         } else {
2336                 leaf->data_offset = 0;
2337                 error = 0;
2338                 ndata = NULL;
2339         }
2340         if (error)
2341                 goto failed;
2342
2343         /*
2344          * Do the insertion.  This can fail with a EDEADLK or EALREADY
2345          */
2346         cursor->flags |= HAMMER_CURSOR_INSERT;
2347         error = hammer_btree_lookup(cursor);
2348         if (error != ENOENT) {
2349                 if (error == 0)
2350                         error = EALREADY;
2351                 goto failed;
2352         }
2353         error = hammer_btree_insert(cursor, leaf, &doprop);
2354
2355         /*
2356          * Cursor is left on current element, we want to skip it now.
2357          * (in case the caller is scanning)
2358          */
2359         cursor->flags |= HAMMER_CURSOR_ATEDISK;
2360         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2361
2362         /*
2363          * If the insertion happens to be creating (and not just replacing)
2364          * an inode we have to track it.
2365          */
2366         if (error == 0 &&
2367             leaf->base.rec_type == HAMMER_RECTYPE_INODE &&
2368             leaf->base.delete_tid == 0) {
2369                 hammer_modify_volume_field(trans, trans->rootvol,
2370                                            vol0_stat_inodes);
2371                 ++trans->hmp->rootvol->ondisk->vol0_stat_inodes;
2372                 hammer_modify_volume_done(trans->rootvol);
2373         }
2374
2375         /*
2376          * vol0_next_tid must track the highest TID stored in the filesystem.
2377          * We do not need to generate undo for this update.
2378          */
2379         high_tid = leaf->base.create_tid;
2380         if (high_tid < leaf->base.delete_tid)
2381                 high_tid = leaf->base.delete_tid;
2382         if (trans->rootvol->ondisk->vol0_next_tid < high_tid) {
2383                 hammer_modify_volume_noundo(trans, trans->rootvol);
2384                 trans->rootvol->ondisk->vol0_next_tid = high_tid;
2385                 hammer_modify_volume_done(trans->rootvol);
2386         }
2387
2388         /*
2389          * WARNING!  cursor's leaf pointer may have changed after
2390          *           do_propagation returns.
2391          */
2392         if (error == 0 && doprop)
2393                 hammer_btree_do_propagation(cursor, leaf);
2394
2395 failed:
2396         /*
2397          * Cleanup
2398          */
2399         if (error && leaf->data_offset) {
2400                 hammer_blockmap_free(trans, leaf->data_offset, leaf->data_len);
2401
2402         }
2403         hammer_sync_unlock(trans);
2404         if (data_buffer)
2405                 hammer_rel_buffer(data_buffer, 0);
2406         return (error);
2407 }
2408
2409 /*
2410  * Delete the B-Tree element at the current cursor and do any necessary
2411  * mirror propagation.
2412  *
2413  * The cursor must be properly positioned for an iteration on return but
2414  * may be pointing at an internal element.
2415  *
2416  * An element can be un-deleted by passing a delete_tid of 0 with
2417  * HAMMER_DELETE_ADJUST.
2418  *
2419  * This function will store the number of bytes deleted in *stat_bytes
2420  * if stat_bytes is not NULL.
2421  */
2422 int
2423 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2424                         hammer_tid_t delete_tid, uint32_t delete_ts,
2425                         int track, int64_t *stat_bytes)
2426 {
2427         struct hammer_btree_leaf_elm save_leaf;
2428         hammer_transaction_t trans;
2429         hammer_btree_leaf_elm_t leaf;
2430         hammer_node_t node;
2431         hammer_btree_elm_t elm;
2432         hammer_off_t data_offset;
2433         int32_t data_len;
2434         int64_t bytes;
2435         int ndelete;
2436         int error;
2437         int icount;
2438         int doprop;
2439
2440         error = hammer_cursor_upgrade(cursor);
2441         if (error)
2442                 return(error);
2443
2444         trans = cursor->trans;
2445         node = cursor->node;
2446         elm = &node->ondisk->elms[cursor->index];
2447         leaf = &elm->leaf;
2448         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2449
2450         hammer_sync_lock_sh(trans);
2451         bytes = 0;
2452         doprop = 0;
2453         icount = 0;
2454
2455         /*
2456          * Adjust the delete_tid.  Update the mirror_tid propagation field
2457          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2458          */
2459         if (delete_flags & HAMMER_DELETE_ADJUST) {
2460                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2461                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2462                                 icount = -1;
2463                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2464                                 icount = 1;
2465                 }
2466
2467                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2468                 elm->leaf.base.delete_tid = delete_tid;
2469                 elm->leaf.delete_ts = delete_ts;
2470                 hammer_modify_node_done(node);
2471
2472                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2473                         hammer_modify_node_field(trans, node, mirror_tid);
2474                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2475                         hammer_modify_node_done(node);
2476                         doprop = 1;
2477                         if (hammer_debug_general & 0x0002) {
2478                                 hdkprintf("propagate %016jx @%016jx\n",
2479                                         (intmax_t)elm->leaf.base.delete_tid,
2480                                         (intmax_t)node->node_offset);
2481                         }
2482                 }
2483
2484                 /*
2485                  * Adjust for the iteration.  We have deleted the current
2486                  * element and want to clear ATEDISK so the iteration does
2487                  * not skip the element after, which now becomes the current
2488                  * element.  This element must be re-tested if doing an
2489                  * iteration, which is handled by the RETEST flag.
2490                  */
2491                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2492                         cursor->flags |= HAMMER_CURSOR_RETEST;
2493                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2494                 }
2495
2496                 /*
2497                  * An on-disk record cannot have the same delete_tid
2498                  * as its create_tid.  In a chain of record updates
2499                  * this could result in a duplicate record.
2500                  */
2501                 KKASSERT(elm->leaf.base.delete_tid !=
2502                          elm->leaf.base.create_tid);
2503         }
2504
2505         /*
2506          * Destroy the B-Tree element if asked (typically if a nohistory
2507          * file or mount, or when called by the pruning code).
2508          *
2509          * Adjust the ATEDISK flag to properly support iterations.
2510          */
2511         if (delete_flags & HAMMER_DELETE_DESTROY) {
2512                 data_offset = elm->leaf.data_offset;
2513                 data_len = elm->leaf.data_len;
2514                 if (doprop) {
2515                         save_leaf = elm->leaf;
2516                         leaf = &save_leaf;
2517                 }
2518                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2519                     elm->leaf.base.delete_tid == 0) {
2520                         icount = -1;
2521                 }
2522
2523                 error = hammer_btree_delete(cursor, &ndelete);
2524                 if (error == 0) {
2525                         /*
2526                          * The deletion moves the next element (if any) to
2527                          * the current element position.  We must clear
2528                          * ATEDISK so this element is not skipped and we
2529                          * must set RETEST to force any iteration to re-test
2530                          * the element.
2531                          */
2532                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2533                                 cursor->flags |= HAMMER_CURSOR_RETEST;
2534                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2535                         }
2536                         bytes += (ndelete * sizeof(struct hammer_node_ondisk));
2537
2538                         switch(HAMMER_ZONE(data_offset)) {
2539                         case HAMMER_ZONE_LARGE_DATA:
2540                         case HAMMER_ZONE_SMALL_DATA:
2541                         case HAMMER_ZONE_META:
2542                                 hammer_blockmap_free(trans,
2543                                                      data_offset, data_len);
2544                                 bytes += data_len;
2545                                 break;
2546                         default:
2547                                 break;
2548                         }
2549                 }
2550         }
2551
2552         /*
2553          * Track inode count and next_tid.  This is used by the mirroring
2554          * and PFS code.  icount can be negative, zero, or positive.
2555          */
2556         if (error == 0 && track) {
2557                 if (icount) {
2558                         hammer_modify_volume_field(trans, trans->rootvol,
2559                                                    vol0_stat_inodes);
2560                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2561                         hammer_modify_volume_done(trans->rootvol);
2562                 }
2563                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2564                         hammer_modify_volume_noundo(trans, trans->rootvol);
2565                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2566                         hammer_modify_volume_done(trans->rootvol);
2567                 }
2568         }
2569
2570         /*
2571          * mirror_tid propagation occurs if the node's mirror_tid had to be
2572          * updated while adjusting the delete_tid.
2573          *
2574          * This occurs when deleting even in nohistory mode, but does not
2575          * occur when pruning an already-deleted node.
2576          *
2577          * cursor->ip is NULL when called from the pruning, mirroring,
2578          * and pfs code.  If non-NULL propagation will be conditionalized
2579          * on whether the PFS is in no-history mode or not.
2580          *
2581          * WARNING: cursor's leaf pointer may have changed after do_propagation
2582          *          returns!
2583          */
2584         if (doprop) {
2585                 if (cursor->ip)
2586                         hammer_btree_do_propagation(cursor, leaf);
2587                 else
2588                         hammer_btree_do_propagation(cursor, leaf);
2589         }
2590         if (stat_bytes)
2591                 *stat_bytes = bytes;
2592         hammer_sync_unlock(trans);
2593         return (error);
2594 }
2595
2596 /*
2597  * Determine whether we can remove a directory.  This routine checks whether
2598  * a directory is empty or not and enforces flush connectivity.
2599  *
2600  * Flush connectivity requires that we block if the target directory is
2601  * currently flushing, otherwise it may not end up in the same flush group.
2602  *
2603  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2604  */
2605 int
2606 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2607 {
2608         struct hammer_cursor cursor;
2609         int error;
2610
2611         /*
2612          * Check directory empty
2613          */
2614         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2615
2616         cursor.key_beg.localization = ip->obj_localization |
2617                                       hammer_dir_localization(ip);
2618         cursor.key_beg.obj_id = ip->obj_id;
2619         cursor.key_beg.create_tid = 0;
2620         cursor.key_beg.delete_tid = 0;
2621         cursor.key_beg.obj_type = 0;
2622         cursor.key_beg.rec_type = HAMMER_RECTYPE_ENTRY_START;
2623         cursor.key_beg.key = HAMMER_MIN_KEY;
2624
2625         cursor.key_end = cursor.key_beg;
2626         cursor.key_end.rec_type = HAMMER_RECTYPE_MAX;
2627         cursor.key_end.key = HAMMER_MAX_KEY;
2628
2629         cursor.asof = ip->obj_asof;
2630         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2631
2632         error = hammer_ip_first(&cursor);
2633         if (error == ENOENT)
2634                 error = 0;
2635         else if (error == 0)
2636                 error = ENOTEMPTY;
2637         hammer_done_cursor(&cursor);
2638         return(error);
2639 }
2640
2641 /*
2642  * Localize the data payload.  Directory entries may need their
2643  * localization adjusted.
2644  */
2645 static
2646 int
2647 hammer_cursor_localize_data(hammer_mount_t hmp, hammer_data_ondisk_t data,
2648                             hammer_btree_leaf_elm_t leaf)
2649 {
2650         uint32_t localization;
2651
2652         if (leaf->base.rec_type == HAMMER_RECTYPE_DIRENTRY) {
2653                 localization = leaf->base.localization &
2654                                HAMMER_LOCALIZE_PSEUDOFS_MASK;
2655                 if (data->entry.localization != localization) {
2656                         data->entry.localization = localization;
2657                         hammer_crc_set_leaf(hmp->version, data, leaf);
2658                 }
2659         }
2660         return(0);
2661 }